Fedelm Advisory

Fedelm Advisory helps NHS planners, R&D teams and analytics groups forecast outcomes, choose the next experiment and act on ranges rather than points. Probabilistic forecasting, Bayesian modelling, predictive modelling and experimental design by Ellen Marsh, with the baselines, calibration and handover that make the work usable.

Forecasts and models verified and checked.

Forecasting and Bayesian models for financial, experimental forecasting and engineering design decisions, built so the results can be checked.

Start a conversation See the packages

Hi, I’m Ellen Marsh.

I build modelling tools for companies so they can plan, predict and allocate costs, resources and money, and check afterwards that the model earned its keep.

Four skills, one standard.

Every engagement starts with the same question: what decision does this need to support? Then I build the simplest model that answers it, name the baseline it had to beat, measure how often its ranges hold, and hand it over with a one-page memo and a monitoring plan.

Bayesian optimisation and experimental design

Which experiment to run next, and whether a model would have chosen better than you did.

  • Bayesian optimisation and active learning: the next experiment chosen by a model
  • Screening campaigns replayed on their own data against random selection
  • Seed size, batch size and refit planning for the next campaign
  • Reference-run design so that batch drift can be removed

Probabilistic forecasting

Ranges that hold as often as they claim, for planning decisions that repeat.

  • Demand and capacity forecasting for health, energy and operations
  • Baselines, rolling-origin backtests and calibration checks
  • Revision-aware evaluation on the data as it was known at the time
  • Scheduled pipelines, public scorecards and handover packs

Predictive modelling and machine learning

Statistical inference and ML focused on questions that change decisions.

  • Statistical analysis and inference
  • Gradient-boosted and classical models, judged side by side
  • Conformal prediction and interval calibration
  • Model review, validation and forecast audits

Bayesian modelling and measurement

Hierarchical models in PyMC for effects across many units, with the checks that make them defensible.

  • Mix, dose-response and price-response models
  • Identifiability audits and prior and posterior predictive checks
  • Calibration against experiments and holdouts
  • Production wrappers with diagnostic gates in CI

What the work lets you do.

Every package turns a guess into a range with a known hit rate, or a search into a sequence of experiments chosen by a model. Both let a team commit less to be sure: fewer runs, fewer agency shifts, less stock held against a number nobody trusts. The figures below are published results and my own public projects. None is a client result.

45 to 65 vs 95 to 105
Fewer experiments to the same answer
Experiments needed to reach a top-1% cell-culture medium when a model chose each batch, against random selection, on three published campaigns replayed on their own measurements. In a Nature benchmark, Bayesian optimisation beat expert chemists on both the number of experiments and the consistency of the outcome.
29 components, 4 rounds
Experiments deployed in small batches
A published active-learning campaign tuned 29 medium components in four rounds of 18 to 19 cultures each, after 232 starting combinations. The best media beat the commercial formulation from round three, with serum an order of magnitude lower. The same loop runs on formulation, fermentation conditions or assay development.
Hashizume, Ozawa and Ying, npj Systems Biology and Applications (2023).
Under 10% error
Staffing planned a week ahead
A 2023 systematic review found every published model of daily emergency-department visits it examined forecast within 10%. One department has run such a forecast weekly to set staff deployment, with 4.8% error on total attendances. Cover bought at short notice is what a missed forecast costs: NHS England expected £2.1 billion of agency spend in 2024/25.
87.5 to 90.3%
Stock, cash and capacity sized to a range
Share of A&E outturns that fell inside my calibrated 90% range at every horizon from one to six months, on a window sealed before the test; the uncalibrated forecast managed 68 to 75%. A range whose hit rate is known lets a planner hold beds, stock or cash against the cost of being wrong on each side, and the same forecast sizes an engineering design margin.
NHS A&E forecasting project, provider level, 19 sealed origins.

Packages

Four packages.

Each is priced by the day or as a fixed fee, and each ends with a one-page memo for the person making the decision and a repository for whoever inherits the model. The colour matches the skill above.

Campaign replay audit

1 to 2 weeks

Fixed fee, agreed in writing

For an R&D team that screens in batches and wants to know whether a model would have chosen better.

  • One-page replay report from your own CSV
  • Planning table for the next campaign
  • What to run in every batch so drift can be removed

Forecast audit

2 to 4 weeks, typically 4 to 8 days

£500 a day

For a forecast you already rely on and nobody has recently checked.

  • Baseline table and calibration plots
  • One-page memo: trust it, for which horizons, with what range
  • A monitoring plan your own team can run

Probabilistic demand forecasting

8 to 12 weeks, two days a week

Quoted against scope

For a planning decision that repeats: beds, staffing, stock, energy.

  • Quantile forecasts scored on rolling-origin backtests
  • A decision table in your units and a scheduled pipeline
  • Handover pack and a one-page memo

Bayesian measurement

Fixed scope or multi-month

Quoted against scope

For questions where one number would mislead and data arrives from many units at once.

  • Hierarchical model in PyMC with its tests
  • Identifiability report and predictive checks
  • Decision output in your units with its uncertainty

Also available: one-hour expert calls and half-day workshops for analytics teams. Two days a week from September 2026, remote from the UK, inside or outside IR35. All packages, with how buying works.

How the method works

Bayesian optimisation, one experiment at a time.

The graphic below is the method behind the experimental-design packages, running live as you scroll. A model with honest uncertainty chooses each experiment, and the campaign ends when the model says further gains are unlikely.

  1. Start with a few experiments

    Three runs, spread across the range. Nobody yet knows where the response peaks.

  2. Fit a model that admits what it does not know

    A Gaussian process gives a best guess (the line) and a range (the band). The band is wide wherever nothing has been measured.

  3. Score every candidate by expected improvement

    The lower strip weighs how good a point could be against how sure the model is. Its peak is the next experiment.

  4. Run it, refit, repeat

    Each result narrows the band and moves the peak. Experiments so far: 3.

  5. Stop when further gains are unlikely

    After nine runs the best result sits on the peak of the true response, which the model never saw. The campaign replay project measures this saving on real data.

Bayesian optimisation on a synthetic response A Gaussian process mean and 95% band over nine observations, with an expected-improvement strip beneath and a marker at the next chosen experiment.

Illustration on a synthetic response, computed in your browser: a Gaussian process with a squared-exponential kernel and expected improvement, refit after every run. Nothing on this graphic is a client result. See the campaign replay project for the method on real campaigns.

Three projects, scored in public.

Each was pre-registered before I looked at results, reports its baselines before its candidates, measures the coverage of its ranges, and ships as a repository that installs and tests in two commands. The number a buyer checks first:

Months ahead 1 2 3 4 5 6
Calibrated forecast, outturns inside its 90% range 87.5% 88.3% 88.7% 89.6% 89.5% 90.3%
The same forecast, uncalibrated 74.9% 72.0% 71.9% 70.8% 69.3% 68.3%
NHS England A&E, provider level, 19 forecast origins sealed before the test and opened once. Registered tolerance 85 to 95% at every horizon. Source: nhs-ae-forecast/results/H-confirmatory/tables/primary.table.csv.
England Type 1 A&E attendances, every published version since 2015 115 published versions of the monthly series drawn as hairlines, with the latest drawn on top. The 2020 collapse is visible in the middle of the series. 2016 2018 2020 2022 2024 2026 Apr 2020: 0.69m Jul 2026: 1.49m

England’s monthly Type 1 A&E attendances, June 2015 to July 2026, drawn once for each of the 115 dates on which NHS England’s published files changed, with the latest version on top. The lines almost coincide: revisions to this series are small, which is one of the things the NHS project measured. Source: NHS England, archived in nhs-ae-forecast.

  • Attendances and emergency admissions for England, one to six months ahead, from every published version of the data since 2015. Pre-registered, sealed, and published live from 31 October 2026 for all 36 integrated care boards, scored against each month’s outturn from 12 November.
    Probabilistic forecasting and predictive modelling: seasonal naive, ETS, STL+ARIMA, LightGBM with conformal calibration, hierarchical PyMC, MinT reconciliation.
    87.5 to 90.3% of outturns inside the 90% range at every horizon; 68 to 75% without calibration
  • Three published cell-culture media campaigns replayed on their own measurements: a Bayesian surrogate refit every five experiments against random selection, plus a planning table for how to seed and batch the next campaign. The tool takes one CSV and returns a one-page report.
    Bayesian modelling and experimental design: PyMC, GP-BO, a replay CLI. Eleven pre-registered hypotheses across five tags, each with its verdict on the page.
    45 to 65 vs 95 to 105 experiments to a top-1% medium, model-guided against random, on the campaigns’ own enriched pools
  • Eight published cell-culture media studies catalogued to one schema, each with a dataset card recording units, design, replicate structure, licence and the loading decisions applied, and a one-command loader.
    Messy-data standardisation, the same method as a published review of more than 100 studies.
    8 studies on one schema with a licence register; six mirrored, two as fetch instructions where the licence forbids it

Let’s talk about your decision.

Send a note describing the decision you’re trying to make, the data that exists, and when you need an answer. I reply within two working days, and within a week you’ll have a written scope, a day count and a fee, or a short note saying it isn’t a fit and who might be.

ellen@fedelm.co.uk

Start a conversation Code on GitHub